AI & Automation

Data Readiness for AI: A Checklist to Run Before Any Project Starts

November 11, 2026 · 4 min read · Etzal Group Team

When an AI project runs late, the model is rarely the reason. The usual cause is discovered in week three: the data is scattered across spreadsheets, half the records are duplicates, and nobody is sure the company is permitted to use them this way. A data readiness check, run before the project begins, turns those surprises into a list.

The checklist below takes a few days for a single workflow. It asks six questions.

1. Where does the data live?

Write down every place the workflow touches: the ERP, the CRM, shared drives, email inboxes, chat exports, paper forms that are scanned later. For each one, note:

  • The system and who administers it.
  • Whether it has an API, a scheduled export, or only a screen.
  • How far back the history goes.

A workflow whose data sits in one system with an API is ready in a way that a workflow spread across five personal spreadsheets is not. Knowing which you have changes the plan and the budget.

2. Who owns it?

For each source, name one person in the business who can answer questions about what the fields mean and who can approve changes. If nobody can explain why two columns both say "customer type", that is a finding. It will slow the project until somebody decides.

3. How good is it?

Do not judge quality from impressions. Pull a random sample of 200 records and count:

  • Completeness. How many required fields are empty?
  • Duplicates. How many customers, products or suppliers appear more than once under slightly different names?
  • Consistency. Are dates, units, currencies and codes written the same way throughout?
  • Accuracy. For twenty records, compare against the real world: call the number, check the document.
  • Freshness. When was each record last updated?

Report the result as percentages. "Twelve percent of customer records have no phone number" is something a team can fix. "The data is messy" is not.

4. Are you allowed to use it?

This question stops more projects than any technical one, and it is best asked early.

  • Does the data include personal information about customers, employees or patients?
  • What did you tell those people when you collected it, and does the new use fit that purpose?
  • Where will the data be processed, and does it leave the country?
  • Will it be sent to a third-party model provider, and what do their terms say about retention and training?
  • Which staff should be able to see which records, and will the AI system respect the same limits?

Indonesia's Personal Data Protection Law and Singapore's PDPA both expect you to have answers to these. Involve whoever handles legal and compliance before data is copied anywhere.

5. Is it in a shape the task can use?

Different tasks need different shapes:

  • A knowledge assistant needs documents that are current, not contradictory, and readable as text. Scanned images need text extraction first.
  • Document extraction needs a set of real example documents with the correct values written down, so results can be checked.
  • Prediction or scoring needs history with the outcome recorded. If you want to predict which leads convert, you need to know which past leads did.
  • An agent that acts needs reliable identifiers, so that "this customer" in one system is the same customer in another.

6. Can it be kept in shape?

A one-time clean-up decays. Ask how new data enters, and whether errors can be prevented at entry: required fields, dropdowns in place of free text, duplicate warnings, and a monthly check of the same quality measures you counted in step three.

Scoring the result

For each workflow you are considering, mark the six questions green, amber or red.

  • Mostly green: the project can start, and data work is a small part of it.
  • Amber on quality or shape: start with a data preparation phase and budget for it explicitly.
  • Red on ownership or permission: resolve it before anything else. No model fixes a missing owner or a missing legal basis.

It is common for the first project on a roadmap to be chosen partly because its data is ready. That is a sound reason. A modest workflow with clean data delivers sooner than an ambitious one that needs a year of preparation.

What to do with the findings

Turn each red and amber item into a task with an owner and a date. Some are quick, such as adding a required field. Some are projects in themselves, such as merging duplicate customer records across two systems. Put them on the same plan as the AI work, because they decide its timing.

A data readiness review is part of every AI Consulting engagement we run, and when the gaps are in the systems themselves, our Enterprise Software team builds the integrations and data fixes. You can write to us with a description of the workflow and its sources.

Summary

Find the data, name its owner, measure its quality on a sample, confirm you may use it, check its shape against the task, and plan how it stays clean. Six questions asked early save months later.

Frequently asked questions

How clean does our data need to be before starting an AI project?

Clean enough for the specific task. A drafting assistant can work with imperfect documents if a person reviews the output. A system that acts on its own, such as automatic invoice posting, needs much higher accuracy in the fields it relies on. Test a sample for the task you have in mind.

Do we need a data warehouse before using AI?

No. Many useful projects read from the systems you already have, through an API or an export. A warehouse helps when you combine several sources for analysis, but it is not a precondition for a first project.

Who should be responsible for data quality?

The team that creates the data, not IT alone. Sales owns customer records, finance owns ledger entries, operations owns stock data. IT provides the tools and access controls. Without a named owner in the business, quality problems return after every clean-up.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

WhatsApp AI Chatbot

We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.